Automating Enterprise Data Engineering with LLMs

The automation of data engineering tasks is invaluable for enterprises to increase efficiency and reduce the manual effort associated with handling large amounts of data. Large Language Models (LLMs) have recently shown promising results in enabling this automation. However, data engineering tasks in real-world enterprise scenarios are often more complex than their typical formulations in the scientific community. In this paper, we study the challenges that arise when automating real-world enterprise data engineering tasks with LLMs. As part of the paper, we perform a case study on the task of matching incoming payments to open invoices, an instance of the entity matching problem. We also release a hand-crafted dataset based on the actual enterprise scenario to enable the research community to study the complexity of such enterprise tasks.

Please check out our paper and cite our work:

@inproceedings{bodensohn2024automating,
	title={Automating Enterprise Data Engineering with {LLM}s},
	author={Jan-Micha Bodensohn and Ulf Brackmann and Liane Vogel and Anupam Sanghi and Carsten Binnig},
	booktitle={NeurIPS 2024 Third Table Representation Learning Workshop},
	year={2024},
	url={https://openreview.net/forum?id=m85fYEJtDc}
}

Setup

Make sure you have Conda installed.

Create a new Conda environment, activate it, and add the project to the Python path:

conda env create -f environment.yml
conda activate llmeval-trl24
export PYTHONPATH=${PYTHONPATH}:./

Payment-to-Invoice Matching Dataset

We release our hand-crafted dataset for the entity matching scenario of matching incoming payments to open invoices. We provide multiple versions of the dataset with different perturbation modes and schema modes:

Perturbation modes:

• single: each payment has zero or one perturbation
• multi: each payment can have multiple perturbations

Schema modes:

• descriptive: the columns have descriptive names (e.g., Client, Billing Number, etc.)
• opaque: the columns have opaque names (e.g., MANDT, VBELN, etc.)
• multi-table: the columns have opaque names and invoices are represented by rows in multiple tables

The dataset is located in data/entity_matching/pay_to_inv/download/<perturbation_mode>/<schema_mode>. Invoices are stored in invoices.csv and payments in payments.csv. Which payments belong to which invoices is stored in matches.csv, which also contains the match categories (1:1, 1:N, N:1) and perturbation categories (e.g., small_deduction).

The dataset generator is implemented in scripts/entity_matching/pay_to_inv/download.py and configurable in config/entity_matching/dataset/pay_to_inv.yaml.

To re-create the dataset, run:

bash scripts/entity_matching/create_dataset.sh

Reproducibility

We provide code to create the dataset and reproduce the results from the paper.

Reproducing the exact results from the paper requires the following artifacts:

• openai_cache.zip the OpenAI API requests and responses, which you must unpack into data/openai_cache

To create the dataset and reproduce the results from the paper, run:

bash reproduce.sh

The results are:

• data/entity_matching/increasing_difficulty.csv Table 1 (F1 scores at increasing difficulties)
• data/entity_matching/precision_recall.pdf Figure 2 (precision and recall for + multi-matches scenario)
• data/entity_matching/error_categories.pdf Figure 3 (F1 scores for typical error categories)